The Limits Tightened and Nobody Sent a Memo

Read the AI forums this fortnight and a single complaint keeps surfacing in different clothes: I am paying the same or more, and getting less. Fewer tokens per prompt. Tighter weekly caps. A model that argues instead of answering. A safety classifier that won’t run an ordinary command. None of it is a scandal. All of it is the accumulated texture of the AI price war as felt from the paying end.

We went looking for the specifics on Hacker News, where the gripes arrive with a username, a timestamp and a permalink you can check yourself. A note on method, because we hold ourselves to it: we wanted Reddit and X in here too, but both were unreachable to us at the time of writing, so everything below is from Hacker News and every quote is verifiable at the links in the Sources list. We quote people sympathetically. The target is the product and the pricing, never the person typing.

The through-line, if you want it in one sentence: the meter is winning, and it is winning quietly.

The subscription that quietly shrank

The most common shape of complaint is not “this is expensive” but “this got smaller without anyone saying so.” On a thread titled “Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive,” vincent_s wrote: “It definitely feels like subscription limits have dropped since GPT-5.6 came out.” The word doing the work is feels — the change is perceptible but unannounced, so users are left reverse-engineering their own allowances.

In the same discussion, esperent put a number on it: “A $200 plan that burns through it’s weekly limit in a day, sometimes in 12 hours, is not sustainable.” And satvikpendem summarised the widely felt asymmetry between vendors: “Codex has much higher, almost unlimited limits, while Claude Code rate limits hourly and weekly much more.” This is the lived experience behind our long-running complaint that rate limits remain genuinely confusing to the people paying for them.

Paying more for the same words

The second theme is subtler and more corrosive: the suspicion that a price cut on paper is not a price cut in use. tedbradley laid out the mechanism with unusual precision: “With those rolling windows and the opaque ‘pricing’ associated with them, an 80% cut to GPT-5.6 Luna might now translate into an 80% cut to using GPT-5.6 Luna with a sub.” A cheaper token is cold comfort if your subscription simply lets you buy fewer of them.

For apatheticonion, coming to an Anthropic shop from DeepSeek, the cost showed up as latency and burn: “It’ll spend 30 minutes thinking and charge like $12. And no token caching, what are you even doing Anthropic? … It’s unusable.” To be fair, he conceded the trade-off first — Sonnet 5 is, he allowed, “a little better on long horizon tasks and headless unsupervised agent workflows” — before the verdict landed; his objection is that the cost and latency make it the wrong tool for the interactive loop he actually works in. And RussianCow, comparing coding subscriptions, reached the conclusion the whole sector keeps circling: “There’s no way their current pricing is sustainable.” If the economics don’t close, the person who eventually pays for that is the subscriber — a theme we picked apart in the hidden cost of tokens.

Capability, quietly clipped

Sometimes what shrinks is not the allowance but the model. Reacting to OpenAI reducing Codex’s context window from 372k to 272k tokens, KronisLV explained why a spec cut lands as a capability cut for real work: for heavy documentation and multi-repo sessions, “limited context size would be a dealbreaker,” and even now, he wrote, “it’s like a slot machine after compacting the context, sometimes steps or other details just evaporate in thin air.”

Others point at opacity dressed as a feature. dudeinhawaii, who uses every major provider daily, singled out one dark-pattern-shaped annoyance: the “Gemini web UI annoying resets to its lowest intelligence which feels scummy and Google is not transparent about what ‘extended thinking’ really is.” When you can’t tell which model you are actually being served, you can’t tell what you are paying for.

The drift you have to babysit

A quieter complaint runs under the loud ones: the tools increasingly need managing to stay on task. vincent_s described a test that is its own small indictment. He stopped “four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it ‘Please check if you’re really going against the actual goal or have you drifted away from that?’ and all four replied something like ‘Yes: I had started to drift’.” You are paying for autonomy and getting something that wanders off unless watched.

esperent has simply built scaffolding around the problem. To keep the agent from exceeding scope overnight, he resorts to “calling the plan a ‘PRD contract’ and a looping message reminding it not to go out of scope.” It works, he says — but note what it is: an experienced user inventing rituals to keep an expensive tool inside the lines. The productivity story sold on the pricing page rarely itemises the supervision tax. It is the daily-driver version of the point we made about the myth of the tenfold developer.

Computer says no

Then there is the refusal tax — the responsible user paying for the vendor’s caution. user43928 described a coding agent that turned an everyday command into a standoff: the classifier “would start to yap about not being able to verify the repository’s privacy settings” on a plain git push, and — the part that stings — “my bug reports went ignored, the issue continued, and I eventually just switched to Full access.” The safety feature trained a careful user to turn safety off.

It is not only chat and code. On the video side, echelon described creative work colliding with trigger-happy moderation: “hair-trigger platform safety checkers … Video models are notoriously bad at shutting down a huge number of requests.” Same tax, different medium. echelon’s fuller point is that this is why serious creative work is fleeing to private GPU clusters — not cost alone, but the refusal to let a legitimate batch of generations run without a moderation layer second-guessing every prompt. When the safest path is to leave the platform, the platform has mispriced safety. We have argued this at length in when AI refuses perfectly normal requests: the cost of a blunt filter is borne by the people using the tool correctly.

The escape hatch everyone names

What makes this fortnight’s mood different from ordinary grumbling is that the alternatives are now credible, and users say so unprompted. eth0up captured the exasperation with a model’s manner rather than its price — our moan of the day, below — while others were already packing. esperent again: “It seems Qwen / Kimi latest are at least equal to GPT 5.5 so as soon as someone else is offering those with a decent rate or subscription, I’m bouncing.” apatheticonion was blunter still: “DeepSeek are in a league of their own.”

The list is specific and it recurs. KronisLV, weighing whether to keep paying the majors, noted the pull of the challengers on exactly the axis that hurts: “Luckily DeepSeek V4 Pro, GLM 5.2 and Kimi K3 don’t seem to have those limits either.” And dudeinhawaii has already voted with his wallet on one incumbent: “I had the Ultra plan and cancelled it once it was apparent they were not improving the agentic experience nor trying to compete.” When the complaint ends in a cancellation and a named replacement, it has stopped being a grumble and started being churn.

The Chinese labs keep coming up not as a geopolitical talking point but as a practical one: cheaper, fewer limits, good enough for the job. That is what a competitive market sounds like from the customer’s chair, and it is the reason the incumbents’ quiet tightening is a risk rather than a free win.

The pattern under the gripes

Step back from the individual annoyances and they rhyme. The recurring categories, roughly in order of how often they surfaced:

• Shrinking allowances — the same price buying fewer tokens or a tighter weekly cap, without an announcement.
• Opaque metering — rolling windows and undisclosed model-swaps that make a bill impossible to predict.
• Capability cuts — smaller context, lossy compaction, a downgraded model served silently.
• Refusal friction — safety classifiers blocking ordinary work, with support that doesn’t answer.
• Credible defection — a named list of cheaper rivals users are ready to switch to.

To be fair to the companies, none of this is evidence of bad faith, and most of these people plainly still like the tools — they are power users, not haters, and several are describing products they use for hours a day. Capacity is genuinely scarce, frontier inference is genuinely expensive, and rolling limits are a defensible way to keep a service up. The complaint is not that limits exist. It is that they move without notice, in a direction that always favours the meter, described in language vague enough that you only discover the new boundary by hitting it.

It is worth saying what this fortnight is not. It is not a claim that the models got worse at everything, or that the companies are acting in bad faith, or that free lunches were owed. Inference genuinely costs money, and someone has to pay for it. The narrower, fairer complaint is about notice and legibility: allowances that move without a changelog, models swapped underneath a familiar name, and “pricing” whose scare-quotes the users are now adding themselves.

If there is one thing to take from a fortnight of other people’s receipts, it is to keep your own. Measure what your subscription actually gives you this week versus last, watch which model you are really being served, and treat the growing list of cheaper rivals not as disloyalty but as leverage. The quiet tightening only works while nobody is counting. These people are counting, out loud, with timestamps — and that is the most pro-consumer thing on the whole thread.